In order to bring you the best possible user experience, this site uses Javascript. If you are seeing this message, it is likely that the Javascript option in your browser is disabled. For optimal viewing of this site, please ensure that Javascript is enabled for your browser.
Did you know that your browser is out of date? To get the best experience using our website we recommend that you upgrade to a newer version. Learn more.

Can CAD prediction be improved with artificial intelligence?

EuroEcho 2019 Congres News

Coronary Artery Disease (Chronic)

Mr_Ross _Upton_2019.jpgMr. Ross Upton, Founder and Chief Executive Officer of Ultromics Ltd, Oxford, UK, which has spun out of the University of Oxford, gave an enlightening Late-Breaking Science presentation yesterday on a new artificial intelligence (AI)-based diagnostic support tool for the prediction of coronary artery disease (CAD).

Mr. Upton explains, “Imaging is central to most diagnostic pathways for CAD, but currently the accuracy of interpretation, for example, of stress echocardiography, is highly variable.”

“Our aim is to automate image analysis and interpretation processes and to create AI-based technology that is at least 80% accurate at predicting CAD—we don’t want to replace clinicians, but we want to help improve accuracy.”

The EVAREST Trial, administered by the Oxford Cardiovascular Clinical Research Facility at the University of Oxford, involves 5,000 participants across 30 UK National Health Service (NHS) sites. Over 100,000 images from EVAREST, including both contrast and non­contrast stress and rest images, were used to construct the automated AI image processing pipeline. The platform was trained to assess the quality of images, identify views, segment the left ventricle (LV), select cycles and frames, quantify novel and standard metrics, and to provide a binary risk prediction for CAD.

“We have made good progress to date,” says Mr. Upton. “The automated image quality assessment, view and contrast identification, and LV segmentation models performed at accuracies of more than 90%.” Furthermore, the platform appeared to be able to predict CAD. In a dataset of 578 patients, a CAD risk prediction model using features extracted from echocardiographic images achieved an average area under the receiver operator characteristic curve of 0.91. An improvement on this performance was observed with the addition of a priori clinical information.

“Existing technology does not have the capacity to correctly deal with frame rate/heart rate issues or to work with contrast images, but this does not appear to be a problem for our platform,” he continues. “However, it is not just the ultimate accuracy that is important—we need to establish generalisability. This technology was built and tested on the EVAREST data set, but we need to know that it works in every instance. We are currently testing the platform further and have ongoing studies in some major centres in the US and UK.”

Mr. Upton describes an additional benefit that has come out of this work. “While developing echocardiography-focused AI techniques robust enough to handle stress imaging, we realised we had generated a powerful AI platform capable of processing any echocardiography image to derive both routine clinical echocardiographic measurements and to discover novel imaging biomarkers, such as ‘rectangularity’ and ‘solidity’. Ongoing explorations of these echocardiographic biomarkers show great potential for the stratification of patients with CAD.”

Mr. Upton concludes, “The team working on this includes medical doctors and echocardiographers as well as scientists who have worked in hospitals themselves. Therefore, everyone is all too aware of the need to create technology that will actually work ‘on the ground’ where it is needed, and we will continue to develop the platform within an even more robust framework.”